425
Peer Effects in the Workplace
†By Thomas Cornelissen, Christian Dustmann, and Uta Schönberg* Existing evidence on peer effects in the productivity of coworkers stems from either laboratory experiments or real-world studies referring to a specific firm or occupation. In this paper, we aim at providing more generalizable results by investigating a large local labor market, with a focus on peer effects in wages rather than productivity. Our estimation strategy—which links the average permanent productivity of workers’ peers to their wages—circumvents the reflection problem and accounts for endogenous sorting of workers into peer groups and firms. On average over all occupations, and in the type of high-skilled occupations investigated in studies on knowledge spillover, we find only small peer effects in wages. In the type of low-skilled occupations analyzed in extant studies on social pressure, in contrast, we find larger peer effects, about one-half the size of those identified in similar studies on productivity. (JEL J24, J31, J41, M12, M54)
The communication and social interaction between coworkers that necessarily occur in the workplace facilitate comparison of individual versus coworker produc-tivity. In this context, workers whose productivity falls behind that of coworkers, or falls short of a social norm, may experience personal feelings of guilt or shame. They may then act on these feelings by increasing their own efforts, a mechanism referred to in the economic literature as “peer pressure.” Social interaction in the workplace may also lead to “knowledge spillover” in which coworkers learn from each other and build up skills that they otherwise would not have. Such productivity enhanc-ing peer effects may exacerbate initial productivity differences between workers and increase long-term inequality when high-quality workers cluster together in the same peer groups. Moreover, while knowledge spillover is an important source of agglomeration economies (e.g., Lucas 1988; Marshall 1890), social pressure further implies that workers respond not only to monetary but also to social incentives, which may alleviate the potential free-rider problem inherent whenever workers work together in a team (Kandel and Lazear 1992).
* Cornelissen: Department of Economics, University of York, Heslington, York YO10 5DD, United Kingdom (e-mail: [email protected]); Dustmann: Department of Economics, University College London, 30 Gordon Street, London WC1H 0AX, United Kingdom, and CReAM (e-mail: [email protected]); Schönberg: Department of Economics, University College London, 30 Gordon Street, London WC1H 0AX, United Kingdom, CReAM, and IAB (e-mail: [email protected]). We are grateful to Josh Kinsler for comments and for sharing program code. Uta Schönberg acknowledges funding by the European Research Council (ERC) Starter grant SPILL. Christian Dustmann acknowledges funding by an ERC Advanced Grant. The authors declare that they have no relevant or material financial interests that relate to the research described in this paper.
† Go to https://doi.org/10.1257/aer.20141300 to visit the article page for additional materials and author
Despite the economic importance of peer effects, empirical evidence on such effects in the workplace is as yet restricted to a handful of studies referring to very specific settings, based on either laboratory experiments or on real-world data from a single firm or occupation. For instance, Mas and Moretti’s (2009) study of one large supermarket chain provides persuasive evidence that workers’ productivity increases when they work alongside more productive coworkers, a finding that they attribute to increased social pressure. Likewise, a controlled laboratory experiment by Falk and Ichino (2006) reveals that students recruited to stuff letters into enve-lopes work faster when they share a room than when they sit alone.1 For peer effects
in the workplace induced by knowledge spillover, however, the evidence is mixed. Whereas Waldinger (2012) finds little evidence for knowledge spillover among scientists in the same department in a university, Azoulay, Graff Zivin, and Wang (2010) and Jackson and Bruegmann (2009) find support for learning from cowork-ers among medical science researchcowork-ers and teachcowork-ers, respectively.2 In a
comprehen-sive meta-analysis of peer effects in coworker productivity covering both high- and low-skilled tasks, Herbst and Mas (2015) report a remarkable similarity between the cross-study average peer effect from laboratory experiments and field studies. Nonetheless, although existing studies provide compelling and clean evidence for the existence (or absence) of peer effects in specific settings, it is unclear to what extent their findings, which are all based on laboratory experiments or real-world studies referring to a specific firm or occupation, apply to the labor market in general.
In this paper, therefore, we go beyond the existing literature to investigate peer effects in the workplace for a representative set of workers, firms, and sectors. Our unique dataset, which encompasses all workers and firms in one large local labor market over nearly two decades, allows us to compare the magnitude of peer effects across detailed sectors. It thus provides a rare opportunity to investigate whether the peer effects uncovered in the literature are confined to the specific firms or sectors studied or whether they carry over to the general labor market, thus shedding light on the external validity of the existing studies. At the same time, our comparison of the magnitude of peer effects across sectors provides new evidence on what drives these effects, whether social pressure or knowledge spillover.
In addition, unlike the extant studies, our analysis focuses on peer effects in wages rather than productivity thereby addressing for the first time whether or not
1 Other papers focusing on social pressure include Kaur, Kremer, and Mullainathan (2010), who report
pro-ductivity spillovers among data entry workers seated next to each other in an Indian company. Similarly, Bandiera, Barankay, and Rasul (2010) find that soft-fruit pickers in one large UK farm are more productive if at least one of their more able friends is present on the same field but less productive if they are the most able among their friends. Peer pressure is also a likely channel in the Chan, Li, and Pierce (2014) finding of peer effects in productivity among salespersons of a department store, and in the general pattern of network effects in the productivity of call center workers found by Lindquist, Sauermann, and Zenou (2015). In work that analyzes regional shirking differen-tials in a large Italian bank, Ichino and Maggi (2000) find that average peer absenteeism has an effect on individual absenteeism. A controlled field experiment by Babcock et al. (2015) suggests that agent awareness of their own efforts’ effect on known peer payoffs creates incentives possibly mediated by a form of social pressure.
2 In related work, Waldinger (2010) shows that faculty quality positively affects doctoral student outcomes,
while Serafinelli (2013) provides evidence that worker mobility from high- to low-wage firms increases the produc-tivity of low-wage firms, which is consistent with knowledge spillover. The findings by Lindquist, Sauermann, and Zenou (2015) and De Grip and Sauermann (2012) suggest knowledge transfer to be a relevant source of produc-tivity spillover when trained and untrained workers interact. Other studies (e.g., Guryan, Kroft, and Notowidigdo 2009; Gould and Winter 2009) analyze such knowledge spillover between teammates in sports. For a nontechnical discussion of evidence and implications of peer effects in coworker productivity, see also Cornelissen (2016).
workers are rewarded for a peer-induced productivity increase through wages. To do so, we first develop a simple theoretical framework in which peer-induced pro-ductivity effects arise because of both social pressure and knowledge spillover and translate into peer-related wage effects even when the firm extracts the entire surplus of the match. The rationale underlying this result is that, if a firm wants to ensure that workers remain with the company and exert profit-maximizing effort, it must compensate them for the extra disutility from exercising additional effort because of knowledge spillover or peer pressure.
In the subsequent empirical analysis, we then estimate the effect of the long-term or predetermined quality of a worker’s current peers—measured by the average wage fixed effect of coworkers in the same occupation and workplace (which we will refer to as “firms” for brevity)—on the current wage, a formulation that directly corresponds to our theoretical model. We implement this approach using an algo-rithm developed by Arcidiacono et al. (2012), which allows simultaneous estimation of both individual and peer group fixed effects. Because we link a worker’s wage to predetermined characteristics (i.e., the peers’ average worker fixed effect) rather than to peer group wages or effort, we avoid Manski’s (1993) reflection problem.
To deal with worker sorting (i.e., the fact that high-quality workers may sort into high-quality peer groups or firms), we extend the worker and firm fixed effects model pioneered by Abowd, Kramarz, and Margolis (1999) and estimated in for instance Card, Heining, and Kline (2013) and condition on an extensive set of fixed effects. First, by including worker fixed effects in our baseline specifica-tion, we account for the potential sorting of high-ability workers into high-ability peer groups. In addition, to account for the potential sorting of high-ability work-ers into firms, occupations, or firm-occupation combinations that pay high wages, we include firm-by-occupation fixed effects. To address the possibility that firms may attract better workers and raise wages at the same time, we further include time-varying firm fixed effects (as well as time-varying occupation fixed effects). As argued in Section IIA, this identification strategy is far tighter than most strategies used to estimate peer effects in other settings.
On average, we find only small, albeit precisely estimated, peer effects in wages: a one standard deviation increase in peer ability increases wages by 0.3 percentage points. Even if peer effects are small on average for a representative set of occupa-tions, they might still be substantial for specific occupations. In fact, the specific occupations and tasks analyzed in the existent studies on peer pressure (i.e., super-market cashiers, data entry workers, envelope stuffers, fruit pickers) are occupations in which there is more opportunity for coworkers to observe each other’s output, a prerequisite for peer pressure buildup. Similarly, the specific occupations and tasks analyzed in the studies on knowledge spillover (i.e., scientists, teachers) are high-skilled and knowledge intensive, making learning from coworkers particularly important.
In a second analytical step, therefore, we restrict our analysis to occupations sim-ilar to those studied in that literature. Nevertheless, in line with Waldinger (2012), in occupations for which we expect knowledge spillover to be important (i.e., occu-pations that are particularly innovative and demand high skills), we again find only small peer effects in wages. On the other hand, in occupations where peer pressure tends to be more important (i.e., where the simple repetitive nature of the tasks
makes output more easily observable to coworkers), we find larger peer effects. In these occupations, a 10 percent increase in peer ability increases wages by 0. 6–0.9 percent. Not only are these findings remarkably robust to a battery of robust-ness checks, but we provide several types of additional evidence for social pressure being their primary source. In comparison, Mas and Moretti (2009) and Falk and Ichino (2006), studying peer effects in productivity (rather than in wages) in one large supermarket chain and in a controlled laboratory experiment, find effects that are about twice as large. This difference may be because of productivity increases not translating one-for-one into wage increases.
Our results are important for several reasons. First, our finding of only small peer effects in wages on average suggests that the larger peer effects established in specific settings in existing studies may not carry over to the labor market in general. Overall, therefore, our results suggest that peer effects do not contribute much to inequality in the economy.3 Second, even though our results suggest that
the findings of earlier studies cannot be extended to the entire labor market, they do indicate that they can be generalized beyond the single firm or single occupa-tion on which they are based. That is, our findings highlight larger peer effects in low-skilled occupations in which coworkers can, because of the repetitive nature of the tasks performed, easily judge each other’s output—which are exactly the type of occupations most often analyzed in earlier studies on peer pressure. Our findings also add to the existing studies by showing that in such situations, peer effects lead not only to productivity spillover but also to wage spillover, as yet an unexplored topic in the literature.
The remainder of the paper is structured as follows. The next section outlines a theoretical framework that links peer effects in productivity engendered by social pressure and knowledge spillover to peer effects in wages. It also clarifies the inter-pretation of the peer effect identified in the empirical analysis. Sections II and III then describe our identification strategy and our data, respectively. Section IV reports our results, and Section V summarizes our findings.
I. Theoretical Framework
To motivate our empirical analysis, we develop a simple principal-agent model of unobserved worker effort in which peer effects in productivity translate into peer effects in wages, and model two channels: knowledge spillover and social pres-sure. Here, we focus on the basics of the model, and delegate details to online Appendix A.
A. Basic Model
Production Function and Knowledge Spillover.—Consider a firm that employs N workers. In the theoretical analysis, we abstract from the endogenous sorting of workers into firms, which our empirical analysis takes into account. We next
3 Our general result of no strong peer effects within firms is in line with a recent paper by Bloom et al. (2015),
suppose that worker i produces individual output f i according to the following pro-duction function:
f i = y i + ε i = a i + e i
(
1 + λ K _ a~i
)
+ ε i ,where y i is the systematic component of worker i’s productive capacity, depend-ing on individual ability a i , individual effort e i , and average peer ability (excluding worker i ) _ a ~i . In this production function, individual effort and peer ability are com-plements, meaning that workers benefit from better peers only if they themselves expend effort. In other words, the return to effort is increasing in peer ability, and the greater this increase, the more important the knowledge spillover captured by the parameter λ K .4 The component ε
i is a random variable reflecting output vari-ation that is beyond the workers’ control and has an expected mean of zero. Firm productivity simply equals the sum of worker outputs. Whereas a worker’s ability is exogenously given and observed by all parties, effort is an endogenous choice vari-able. As is standard in the principal agent literature, we assume that the firm cannot separately observe either worker effort e i or random productivity shocks ε i .
Cost of Effort and Social Pressure.—Because exerting effort is costly to the worker, we assume that in the absence of peer pressure, the cost of effort function is quadratic in effort: C ( e i ) = k e i2 . As in Barron and Gjerde (1997), Kandel and Lazear
(1992), and Mas and Moretti (2009), we introduce peer pressure by augmenting the individual cost of effort function C( · ) with a social “peer pressure” function P( · ) that depends on individual effort e i and average peer output _ f ~i (excluding worker i ). We propose a particularly simple functional form for the peer pressure function: P ( e i , _ f ~i ) = λ P (m − e
i ) _
f
~i , where λ P and m can be thought of as both the “strength”
and the “pain” from peer pressure. The total disutility associated with effort thus becomes c i = C
(
e i)
+ P ( e i , _ f ~i ) = k e i2 + λ P(
m − e i)
_ f ~i .Although the exact expressions derived in this section depend on the specific func-tional form for the total disutility associated with effort, our general argument does not. This peer pressure function implies that the marginal cost of worker effort is declining in peer output (i.e., ∂
2 P ( e i , _ f ~i ) _______ ∂ e i ∂ _ f ~i = − λ P < 0
) . This condition implies that it is less costly to exert an additional unit of effort when the quality of one’s
4 It should be noted that, just as in extant studies, this formulation abstracts from the dynamic implications of
knowledge spillover, and is best interpreted as one of contemporaneous knowledge spillover through assistance and cooperation between workers on the job. The underlying rationale is that workers with better peers are more pro-ductive on the job because they receive more helpful advice from their coworkers than if they were in a low-quality peer group. Moreover, although a high-quality coworker may still boost a worker’s productivity even when the two are no longer working together, one would still expect current peers to be more important than past peers. In addition, this specification assumes that own effort and time is required to “unlock” the potential of one’s peers’ ability. This assumption of complementarity between knowledge spillover and effort provision is one of the drivers of why knowledge spillover translates into wages in our model: workers exposed to better peers exert higher effort, for which they have to be compensated in terms of higher wages.
peers is high than when it is low. We further assume that, like Barron and Gjerde (1997), workers dislike working in a high-pressure environment—which amounts to imposing a lower bound on the parameter (capturing the “pain from peers”) in the peer pressure function P( · ) (see online Appendix A.1 for details).
Wage Contracts and Worker Preferences.—Firms choose a wage contract that provides their workers with the proper incentives to exert effort. Because the firm cannot disentangle e i and εi, however, it cannot contract a worker’s effort directly but must instead contract output f i . As is typical in the related literature, we restrict the analysis to linear wage contracts. Contrary to the standard principal agent model, we assume that not only firms but also workers are risk neutral, an assump-tion that simplifies our analysis without being a necessary condiassump-tion for our general argument.
B. The Worker’s Maximization Problem
Because of risk neutrality, workers maximize their expected wage minus the combined cost of effort. As shown in online Appendix A.2, this leads to the first order condition (1) e i = λ P ___ 2k _ e ~i + b ___2k + λ P + b λ K _______ 2k _ a ~i for i = 1, … , N,
where b denotes the slope of the wage contract with respect to worker output. This first order condition not only highlights that equilibrium effort is increasing in peer ability (see last term), either because of peer pressure λ P or knowledge spillover λ K , but also that peer pressure ( λ P > 0) leads to a social multiplier effect whereby the more effort exerted by peers, the more effort exerted by the worker ( e i is increasing in _ e ~i ).
C. The Firm’s Optimization Problem
Firms choose the intercept and slope (or incentive) parameter of the wage con-tract by maximizing expected profits, EP = ∑ i E [ f i − w i ] , taking into account the workers’ optimal effort levels e i⁎ , subject to the participation constraint that work-ers receive a utility that is at least as high as the outside option v( a i ) : E U i ≥ v( a i ) . Assuming that the participation constraint holds with equality so that the firm pushes the worker’s wage to her reservation utility, the firm ultimately rewards the worker for the outside option v
(
a i)
, the cost of effort C(
e i⁎)
, and the disutility from peer pressure P ( e i⁎ , _ y ~i ) :(2) E w i = v
(
a i)
+ C(
e i⁎)
+ P ( e i⁎ , _ y ~i ) .We can then derive the firm’s first order condition and an expression for the slope b ∗ of the optimal wage contract as detailed in online Appendix A.3.
D. The Effect of Peer Quality on Wages
We obtain the average effect of peer ability on wages—our parameter of interest in the empirical analysis—by differentiating equation (2) and taking averages: (3) 1 __ N
∑
i dE w i ____ d _ a ~i = __N 1∑
i b ⁎ ∂ fi ___ ∂ e i d e i⁎ ____ d _ a ~i Wage response to own effort increase + 1 __ N∑
i ∂ P ( e i , _ y ~i ) ________ ∂ _ y ~i d _ y ~i ____ d _ a ~i Wage response to disutility from social pressure ,where all terms are evaluated at optimal effort levels and at the optimal b . The first term captures the wage response to the increase in workers’ own effort and consists of three parts which are all positive: the slope of the wage contract, b ⁎ , the marginal effect of effort on productivity, __∂ f i
∂ e i , and the effect of peer ability on equilibrium
effort, ___d e i⁎
d _ a ~i (see equation (1)). The second term is likewise positive and captures that
higher peer ability is associated with higher peer output ( ___d _ y ~i
d _ a ~i > 0) , which causes
additional “pain” from peer pressure
(
∂ P ______(∂ _ y e i , _ y ~i )~i > 0
)
for which the worker has tobe compensated. Our model thus predicts that the average effect of peer ability on wages will be unambiguously positive—because better peers induce the worker to exert more effort and to work under pressure, for which the firm has to compensate the worker.
II. Empirical Implementation
In our empirical analysis, we seek to estimate the average effect of peer ability on wages as derived in equation (3). While in the model above we abstract from worker sorting, our empirical analysis needs to take account of nonrandom allocation of workers to firms and unobserved background characteristics. As we describe in detail below, to deal with worker sorting, our baseline empirical strategy extends the worker and firm fixed effects model pioneered by Abowd, Kramarz, and Margolis (1999) and conditions on an extensive set of fixed effects. We define a worker’s peer group as all workers working in the same ( three-digit) occupation and in the same firm in period t (see Section IIIB for a detailed discussion of the peer group definition).
A. Baseline Specification and Identification First, we estimate the following baseline wage equation:
(4) ln w iojt = a i + γ _ a ~i,ojt + x ′ iojt β + ω ot + δ jt + θ oj + v iojt .
Here ln w iojt is the individual log real wage, x iojt is a vector of time-varying char-acteristics with an associated coefficient vector β, i indexes workers, o indexes occupations, j indexes workplaces or production sites (which we label “firms” for simplicity), t indexes time periods, and a i is a worker fixed effect. The term _ a ~i,ojt
corresponds to _ a ~i in the theoretical model, and is the average worker fixed effect in the peer group, computed by excluding individual i. The coefficient γ is the param-eter of interest and measures (a positive monotone transformation of) the spillover effect in wages ( __N1 ∑ i ___dEd _ a w ~ii in equation (3)) which embodies not only the direct effect of peer ability on wages, holding peer effort constant, but also the social mul-tiplier effect arising from workers’ effort reactions in response to increases in the current effort of their peers.5 Identifying this reduced-form or total effect of peers’
long-term productivity on wages requires conditioning on _ a ~i,ojt as a measure of the peers’ long-term productivity (a predetermined characteristic), but not conditioning on contemporaneous peer effort or productivity (or as a proxy thereof, peers’ current wages).6
Nonetheless, identifying the causal peer effect γ is challenging because of con-founding factors such as shared background characteristics. Here, we first discuss the conditions required for a causal interpretation of the peer effect γ assuming that a i and _ a ~i,ojt in equation (4) are observed. Then, in Section IIC, we outline the issues arising from the fact that a i and _ a ~i,ojt are unobserved and must be estimated. While we are confident that our estimation strategy results in unbiased estimates of the effect of peer quality on wages, we will argue that any possible remaining bias is likely to be upward, so that our estimates are upper bounds of peer effects.
Peer quality may affect a worker’s wage simply because high-quality work-ers sort into high-quality peer groups or high-quality firms, leading to a spurious correlation between peer quality and wages. Our estimation strategy accounts for the endogenous sorting of workers into peer groups or firms by including control variables and multiple fixed effects. First, because our baseline specification in equation (4) includes worker fixed effects, it accounts for the potential sorting of high-ability workers into high-ability peer groups.7 Second, time-varying
occupa-tion effects ω ot are included to capture diverging time trends in occupational pay differentials. Moreover, our inclusion of time-varying firm fixed effects δ jt controls for shocks that are firm specific. Finally, by controlling for firm specific occupation effects θ oj , we allow for the possibility that a firm may pay specific occupations rel-atively well (or badly) compared to the market.
Estimation of γ in equation (4) exploits two sources of variation: changes in peer quality for workers who switch peer groups (after having controlled for the accom-panying changes in firm and occupation specific fixed effects), and changes in peer quality for workers who remain with their peer group, induced by other workers joining or leaving the peer group. Focusing on the latter source of variation, our identification strategy can essentially be understood as a difference-in-differences estimator. To see this, denote by a ̃ iojt the peer group quality purged from effects of observables ( x iojt ) and occupation-specific shocks common to all firms in the
5 It should be noted that the theoretical wage equation in (3) refers to wage levels, whereas the empirical wage
equation in (4) is estimated in logs. Since the logarithm is a positive monotone transformation, the key prediction of our model carries over to log wages: In the presence of knowledge spillover or peer pressure, both wage levels and log wages are increasing in peer ability (i.e., γ > 0 ).
6 Therefore, there is no reflection problem in estimating the peer effect γ in equation (4) (Manski 1993). 7 Since x
iojt in equation (4) includes quadratics in age and firm tenure, the worker fixed effects are net of age and job tenure, and the average peer fixed effect does not capture effects of peer age or peer job tenure.
economy ( ω ot ) .8 Suppose further for simplicity that all firms consist of two
occu-pations only, denoted by o and o′. First differencing of equation (4) for peer group stayers eliminates the time-constant worker and firm-occupation fixed effects a i and θ oj —and more generally any time-constant effects such as a match-specific effects m ioj —but does not remove the firm-specific shock common to all occupa-tions in the firm ( Δ δ jt ) : Δln w iojt = γΔ a ̃ iojt + Δ δ jt + Δ v iojt . This effect can be eliminated through differencing a second time, between occupations o and o′ in the same firm that experienced different changes in peer quality:
(5) Δ ‾ ln w ojts ⏟ first difference − Δ ‾ ln w os ′ jt second difference = γ
(
Δ a ̃ – ojts − Δ a ̃ – so ′jt)
+(
Δ _ v sojt − Δ _ v os ′jt)
.This firm-level regression consistently estimates γ provided that cov (Δ a ̃ – ojts − Δ a ̃ – o′jt s , Δ v – ojt s − Δ v – o′jt s
) = 0. This condition says that peer group stayers in both occupations in the firm experience the same time shock, which corre-sponds to the standard common time trend assumption in difference-in-differences estimation.
Our identification assumptions are considerably weaker than the assumptions typ-ically invoked in the education literature, which seek to identify exogenous spillover effects (e.g., the impact of the share of girls, Blacks, immigrants, or grade repeaters on individual performance). The most common approach in these studies—measur-ing grade-level peer characteristics and exploitstudies—measur-ing within-school variation over time (e.g., Gould, Lavy, and Paserman 2009; Hanushek et al. 2003; Hoxby 2000; Lavy and Schlosser 2011; Lavy, Paserman, and Schlosser 2012)—does not allow for the possibility that the average quality of students (in our case, workers) in the school (firm) changes over time, or that the school’s effect on student performance (wages) may vary over time. Other research employs an alternative approach: measuring peer characteristics at the classroom level and exploiting within-school grade-year variation (e.g., Ammermueller and Pischke 2009; Betts and Zau 2004; McEwan 2003; Vigdor and Nechyba 2007). This technique, however, requires random assign-ment of students into classrooms within the school (equivalent to occupations within a firm), thereby ruling out within-school student tracking. Our analysis, in contrast, can account for nonrandom selection into occupations within firms by including firm-specific occupation effects.
B. Within–Peer Group Estimator
One remaining concern may be the possible presence of time-varying peer group-specific wage shocks that are correlated with shocks to peer group quality, violating the common time trend assumption highlighted above. The existence of such shocks is likely to lead to a bias. For example, a firm may adopt a new tech-nology specific to one occupation only, simultaneously raising wages and worker
8 That is, a ̃
iojt is the residual from the regression of _
a
quality in that occupation relative to other occupations in the firm, implying that cov (Δ a ̃ – ojts − Δ a ̃ – o′jt s , Δ v – ojt s − Δ v – o′jt s ) > 0.
One way to deal with this problem is to condition on the full set of time-varying peer group fixed effects p ojt . Note that the parameter γ remains identified because focal worker i is excluded from the average peer group quality. As a result, the average peer group quality of the same group of workers differs for each worker, and a – ~i,ojt differs for each worker within a peer group at any given point in time. Using only within–peer group variation for identification yields the following esti-mation equation:9
(6) ln w iojt = a i + γ a – ~i,ojt + x iojtT β + p
ojt + ε iojt .
This within–peer group estimator, however, although it effectively deals with unob-served time-varying peer group characteristics, employs only limited and specific variation in _ a ~i,ojt . That is, as shown in online Appendix B, the spillover effect in equation (6) is identified only if peer groups vary in size. The advantage of being able to control for time-varying shocks to the peer group is thus countered by the disadvantage that only one particular type of variation is used to identify the effect. The within–peer group estimator in equation (6), therefore, serves as a robustness check only, rather than as our main specification.
C. Estimation
Whereas our discussion so far assumes that the individual and average worker fixed effects a i and _ a ~i,ojt are observed, they are in fact unobserved and must be estimated. The multiplication of _ a ~i,ojt and γ , both parameters to be estimated, turns equations (4) and (6) into nonlinear least squares problems. Because the fixed effects are high dimensional (i.e., we have approximately 600,000 firm years, 200,000 occupation-firm combinations, and 2,100,000 workers), using standard nonlinear least squares routines to solve the problem is infeasible. Rather, we adopt the alternative estimation procedure suggested by Arcidiacono et al. (2012), which is detailed in online Appendix C. An appealing characteristic of this estimation pro-cedure is that the nonlinear least squares estimator for γ is consistent as the sample size grows in panels with a limited number of time periods , even though the indi-vidual worker fixed effects a i are generally inconsistent in this situation. This result, however, requires further assumptions in addition to those we discussed above for the case when a i and _ a ~i,ojt are observed (see Theorem 1 in Arcidiacono et al. 2012). Most importantly, the error terms between any two observations ( v iojt in our equation (4) baseline specification and ε iojt in our equation (6) within–peer group estimator) must be uncorrelated. This assumption not only rules out serially cor-related wage shocks, but also, in our baseline regression, any wage shocks common to the peer group, even those uncorrelated with peer group quality. This additional assumption is necessary for consistent estimation when a i and _ a ~i,ojt are unobserved because peer group-specific wage shocks affect not only peer group member wages
9 Because the fixed effects δ
jt , ω ot , and θ oj do not vary within peer groups at any given point in time, they are absorbed by p ojt .
but also estimated fixed effects in panels with a short T. Any such impact could lead to a spurious correlation between individual wages and the estimated worker fixed effects in the peer group even when the peer group-specific wage shocks are uncor-related with the true worker fixed effects in the peer group.
Results from a Monte Carlo study discussed in online Appendix F.1 show that while serial correlation of a plausible magnitude hardly biases peer effect estimates in our application, time-varying random peer group level shocks are likely to lead to an upward bias in panels with short T. However, under realistic assumptions the bias is not large enough to spuriously generate the level of peer effects that we find in low-skilled repetitive occupations. Moreover, results from the Monte Carlo study confirm that even if time-varying peer group shocks were present, the within–peer group estimator of equation (6) deals directly with the bias problem—as it condi-tions on peer-group level wage shocks.
III. Data
Our dataset comes from over three decades of German social security records that cover every man and woman in the system observed on June 30 of each year. This dataset is particularly suited for the analytical purpose because it includes identifiers for single production sites or workplaces (referred to as “firms” for simplicity), as well as detailed occupational codes that distinguish 331 occupations. Such detail allows us to define peer groups of coworkers in the same firm who are likely to inter-act. We can also observe all workers in each firm, which allows precise calculation of the average peer group characteristics and ensures that our findings are represen-tative of both the firms and the workers.
A. Sample Selection
Focusing on the years 1989–2005, we select all workers aged between 16 and 65 in one large metropolitan labor market, the city of Munich and its surrounding dis-tricts.10 Because most workers who change jobs remain in their local labor market,
concentrating on one large metropolitan labor market rather than a random sample of workers ensures that our sample captures most worker mobility between firms, which is important for our identification strategy of estimating firm and worker fixed effects. Since only the fixed effects within a group of firms connected by worker mobility are identified relative to each other, we restrict our sample to the biggest connected mobility group (which makes up 99.5 percent of the initial sample; see Section IIIE for more details).11
Because the wages of part-time workers and apprentices cannot be meaningfully compared to those of regular full-time workers, we base our estimations on full-time 10 We focus on the large metropolitan labor market rather than Germany as a whole in order to reduce the
computational burden, which is far higher than in conventional linear worker and firm fixed effects models (as in, e.g., Card, Heining, and Kline 2013), due to the inclusion of average peer quality in addition to firm-by-time and firm-by occupation-effects. Robustness checks we provide below and comparisons between the Munich area and Germany as a whole discussed in online Appendix F.2 suggest that results for the whole of Germany would not be very different.
11 Two firms are directly connected if worker mobility is observed between them in any sample period. A
workers not in apprenticeship. Additionally, to ensure that every worker is matched with at least one peer, we drop peer groups ( firm-occupation-year combinations) with only one worker.
B. Definition of the Peer Group
We define the worker’s peer group as all workers employed in the same firm and the same three-digit occupation, the smallest occupation level available in the social security data. These include detailed occupational definitions such as bricklayers, florists, plumbers, pharmacists, and high school teachers. Defining the peer group at the three-digit (as opposed to the one- or two-digit) occupation level not only ensures that workers in the same peer group are likely to interact with each other, a prerequisite for knowledge spillover, but also that workers in the same peer group perform similar tasks and are thus likely to judge each other’s output, a prerequisite for peer pressure buildup. Occupations at the two-digit level, in contrast, often lump together rather different occupations. For example, the three-digit occupation of a cashier is part of the same two-digit occupation as accountants and computer and data processing consultants. In online Appendix D, we show that defining the peer group as too large or too small is likely to lead to attenuation bias of the true peer effect. However, the robustness and placebo tests in Tables 6 and 7 suggest that our peer group definition at the three-digit occupation level is the most plausible.
C. Isolating Occupations with High Levels of Peer Pressure and Knowledge Spillover
One important precondition for the buildup of peer pressure is that workers can mutually observe and judge each other’s output, an evaluation facilitated when tasks are relatively simple and standardized but more difficult when job duties are diverse and complex. To identify occupations characterized by more standardized tasks, for which we expect peer pressure to be important, we rely on a further data source, the 1991/1992 wave of the Qualification and Career Survey (see Gathmann and Schönberg 2010, for a detailed description). In addition to detailed questions on task usage, respondents are asked how frequently they perform repetitive tasks and tasks that are predefined in detail. From the answers, we generate a combined score on which to rank occupations. We then choose the set of occupations with the highest incidence of repetitive and predefined tasks, which encompasses 5 percent of the workers in our sample (see column 1 of online Appendix Table F.3 for a full list of the occupations in this group). This group of most repetitive occupations includes agricultural workers, the subject of the Bandiera, Barankay, and Rasul (2010) study, and cashiers, the focus of the Mas and Moretti (2009) study. The remaining occupa-tions are mostly low-skilled manual occupaoccupa-tions, such as unskilled laborers, pack-agers, or metal workers.
For robustness, we also estimate peer effects for the exact same occupations as in the extant studies using real-world data—that is, cashiers (Mas and Moretti 2009), agricultural helpers (Bandiera, Barankay, and Rasul 2010), and data entry workers (Kaur, Kremer, and Mullainathan 2010)—as well as for a handpicked set of low-skilled occupations in which, after initial induction, on-the-job learning is
limited. This subgroup, which includes waiters, cashiers, agricultural helpers, vehi-cle vehi-cleaners, and packagers among others, makes up 14 percent of the total sample (see column 2 of online Appendix Table F.3 for a full list). Unlike the 5 percent most repetitive occupations, this group excludes specialized skill craft occupations in which learning may be important, such as ceramic workers or pattern makers.
To isolate occupations in which we expect high knowledge spillover, we select the 10 percent most skilled occupations in terms of workers’ educational attain-ment (average share of university graduates), which includes not only the scien-tists, academics, and teachers used in previous studies, but also professionals such as architects and physicians. As a robustness check, we also construct a combined index based on two additional items in the Qualification and Career Survey: whether individuals need to learn new tasks and think anew, and whether they need to exper-iment and try out new ideas. From this index, we derive the 10 percent of occupa-tions with the highest scores, which again includes scientists and academics but also musicians and IT specialists. We further handpick a group of occupations that appear very knowledge intensive, including doctors, lawyers, scientists, teachers, and academics (see columns 3 to 5 of online Appendix Table F.3 for a full list of occupations in these three groups).
It should be noted that when focusing on occupational subgroups, we still estimate the model on the full sample and allow the peer effect to differ for both the respective subgroups and the remaining occupations. Doing so ensures that we use all informa-tion available for firms and workers, which makes the estimated firm-year and worker fixed effects—and hence the measure for average peer quality—more reliable.
D. Wage Censoring
As is common in social security data, wages in our database are right censored at the social security contribution ceiling. Such censoring, although it affects only 0.7 percent of the wage observations in the 5 percent most repetitive occupations, is high in occupations with high expected knowledge spillover. We therefore impute top-coded wages using a procedure similar to that employed by Dustmann, Ludsteck, and Schönberg (2009) and Card, Heining, and Kline (2013) (see online Appendix E for details). Whether or not we impute wages, however, our results remain similar even in the high-skilled occupations with high censoring. This finding is not surpris-ing given that censorsurpris-ing generally causes the distributions of both worker fixed effects and average peer quality to be compressed in the same way as the dependent variable, meaning that censoring need not lead to a large bias in the estimated peer effect.
E. Descriptive Statistics
In Table 1, we compare the 5 percent most repetitive occupations, in which we expect particularly high peer pressure, and the 10 percent most skilled occupations, in which we expect high knowledge spillover, against all occupations in our sample. Clearly, the 5 percent most repetitive occupations are low-skilled occupations: nearly half (47 percent) the workers have no post-secondary education (compared to 17 per-cent in the full sample and 4 perper-cent in the skilled occupations sample) and virtually no worker has graduated from a college or university (compared to 18 percent in the
full sample and 80 percent in the skilled occupations sample). As expected, the learn-ing content in the 5 percent most repetitive occupations is low, while it is high in the 10 percent most skilled occupations, as implied by responses to whether individuals need to learn new tasks or to experiment with new ideas. The median peer group size of three or four workers per peer group is similar in all three samples. Not surpris-ingly, peer group size is heavily skewed, with the mean peer group size exceeding the median peer group size by a factor of about 3–4 in the three samples.
To identify peer effects in wages, individual wages must be flexible enough to react to peer quality induced changes in productivity. The evidence presented in Figure 1 and the bottom half of Table 1 illustrates that despite relatively high col-lective bargaining coverage rates in Germany, there is substantial wage variation across coworkers in the same occupation and firm.12 First, the within–peer group
standard deviation of the log wage residuals (obtained from a regression of log wages on quadratics in age and firm tenure and aggregate time trends) accounts for a considerable share of the overall standard deviation: about half in the full sample 12 Collective bargaining agreements in Germany allow for substantial real wage flexibility. First, nothing
pre-vents firms from paying wages above the level stipulated by collective agreements, or to pay extra bonuses based on performance. About 90 percent of workers receive some form of wage supplement on top of their wage base (own calculations based on the German Socio-Economic Panel 1994–2006). Second, wages are usually tied to job titles, not to occupations. Hence within occupations in the same firm, there can be different ranks of job titles into which workers can be promoted based on their productivity. Third, collectively bargained wage floors are agreed in nominal terms, which allows for real wage cuts by freezing nominal wages.
Table 1—Skill Content, Peer Group Size, and Wage Flexibility for Different Occupational Groups
All occupations 5 percent most repetitive occupations 10 percent most skilled occupations Skill content
Share without postsecondary education 0.17 0.47 0.04
Share with university degree 0.18 0.01 0.80
To what extent does the following occur in your daily work? (0 = never, … , 4 = all the time)
Need to learn new tasks and think anew 2.25 1.36 2.98 Need to experiment and try out new ideas 1.80 0.96 2.56 Peer group size
Median 3 4 3
Mean 9.3 12.3 13.1
Wage flexibility
Standard deviation of log real wage (imputed) 0.472 0.326 0.371 Standard deviation of log real wage residuala 0.377 0.308 0.365 Within–peer group standard deviation of log real wage residuala 0.243 0.200 0.269
Probability of >5 percent real wage cut (peer group stayers) 0.088 0.130 0.034
Average wage growth 0.022 0.016 0.023
Notes: The table compares all occupations (N = 12,832,842 worker-year observations) with the 5 percent most repetitive occupations (N = 681,391) and the 10 percent most skilled occupations (N = 1,309,070). See online Appendix Table F.3 for a full list of occupations, and Section IIIC of the text for the definition of repetitive and skilled occupations.
aResidual from a log-wage regression, after controlling for aggregate time effects, education, and quadratics in
firm tenure and age.
Sources: German social security data, one large local labor market, 1989–2005, combined with information from Qualification and Career Survey 1991/1992.
(0.24 versus 0.47), about two-thirds in the 5 percent most repetitive occupations sample (0.20 versus 0.33), and about three quarters in the 10 percent most skilled occupations sample (0.27 versus 0.37). Second, real wages are downwardly flexi-ble: about 9 percent of peer-group stayers in the full sample, 3 percent in the skilled occupations sample, and 13 percent in the repetitive occupations sample experience a real wage cut from one year to another of at least 5 percent. Third, as the figures in the last row of Table 1 illustrate, average real wage growth over our sample period was positive and in the order of 2 percent per year, implying that decreases in pro-ductivity can be accommodated by raising wages more slowly rather than actually cutting nominal or real wages.
We provide additional information on the structure of our sample in Table 2. Our overall sample consists of 2,115,544 workers, 89,581 firms, and 1,387,216 peer groups; workers are observed on average for 6.1 time periods and there are 2.3 peer groups on average per firm and year. Separately identifying worker, firm-occupation and firm-time fixed effects requires worker mobility between occupations and firms. In our sample, workers have on average worked for 1.6 firms and in 1.4 differ-ent occupations. This amount of mobility is sufficidiffer-ent to iddiffer-entify firm-year and firm-occupation fixed effects for nearly the entire sample: the biggest connected groups for firm-time effects and firm-occupation fixed effects contain 99.4 percent and 98.3 percent of the original observations, respectively, compared to 99.5 per-cent for the more standard firm fixed effects.13
13 Note that all firm stayers are “movers” between firm-time units so that it is not surprising that the connected
group is nearly as large for firm-by-time fixed effects as for the firm fixed effects. Further note that unlike standard 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 All
occupations 5% most repetitiveoccupations 10% most skilledoccupations
Standard deviation of log real wage
Standard deviation of residualized log real wages
Within–peer group standard deviation of residualized log real wages
Figure 1. Variability of Wages Across and Within Peer Groups
Notes: The figure compares all occupations (N = 12,832,842), the 5 percent most repetitive occupations (N = 681,391), and the 10 percent most skilled occupations (N = 1,309,070) in terms of the variability of wages. Residualized wages are computed from a log-wage regression controlling for aggregate time effects, education, and quadratics in firm tenure and age. Source: German social security data, one large local labor market, 1989–2005
In our baseline specification based on equation (4), the standard deviation of the estimated worker fixed effects for the full sample ( a i in equation (4)) is 0.32 or 70 percent of the overall standard deviation of log wages. The average worker fixed effects in the peer group (excluding the focal worker _ a ~i,ojt in equation (4)) has a standard deviation of 0.24, which is about 50 percent of the overall standard devia-tion of the log wage.
As explained in Section IIA, our baseline specification identifies the causal effect of peers on wages by exploiting two main sources of variation in peer quality: changes to the peer group make-up as workers join and leave the group, and moves to new peer groups by the focal worker. In Figure 2, we plot the kernel density esti-mates of the change in a worker’s average peer quality from one year to the next separately for those who remain in the peer group (stayers) and those who leave (movers). Not surprisingly, the standard deviation of the change in average peer quality is more than three times as high for peer group movers than for peer group stayers (0.20 versus 0.06; see also Table 2). Yet even for workers who remain in their peer group, there is considerable variation in average peer quality from one year to the next, corresponding to roughly 20 percent of the overall variation in average peer quality. As expected, for peer group stayers, the kernel density has a mass point at zero, corresponding to stayers in peer groups that no worker joins or leaves. Nonetheless, peer groups without turnover are rare. In our sample, 90 percent of observations are in peer groups with at least some worker turnover. At 20 percent,
firm-fixed effects, the firm-occupation fixed effects are identified not only through worker mobility across firms, but also through worker mobility between occupations within firms.
Table 2—Structure of Sample Panel structure
(i) Number of workers 2,115,544
(ii) Number of firms 89,581
(iii) Number of peer groups (occupations within firm years) 1,387,216
(iv) Average number of time periods per worker 6.07
(v) Number of peer groups per firm year 2.30
(vi) Average number of employers per worker 1.60
(vii) Average number of occupations per worker 1.40
(viii) Share of mobility group with identified firm fixed effects 0.995 (ix) Share of mobility group with identified firm-time fixed effects 0.994 (x) Share of mobility group with identified firm-occupation fixed effects 0.983 Variation in wages, peer quality, and worker turnover
(xi) Standard deviation worker fixed effect 0.32
(xii) Standard deviation average peer fixed effect 0.24 (xiii) Standard deviation change of average peer fixed effect from t − 1 to t 0.09 (xiv) Standard deviation change of average peer fixed effect from t − 1 to t − Movers 0.20 (xv) Standard deviation change of average peer fixed effect from t − 1 to t − Stayers 0.06 (xvi) Share of worker-year observations in peer groups with turnover 0.90 (xvii) Average share of workers replaced by turnover 0.20 (xviii) Correlation worker fixed effect/average peer fixed effect 0.64 (xix) Correlation worker fixed effect/ firm-time effect 0.14 (xx) Correlation worker fixed effect/ firm-occupation effect 0.16 Notes: The table shows descriptive statistics describing the panel structure of the dataset, as well as the variation in wages, peer quality and worker turnover which we exploit in subsequent estimations. N = 12,832,842.
the average peer group turnover in our sample, computed as 0.5 times the number of workers who join or leave divided by peer group size, is quite large and implies that nearly 20 percent of workers in the peer group are replaced every year.
At the bottom of the table, we report correlation coefficients between the various fixed effects in equation (4). In our sample, the individual worker fixed effect and the average fixed effect of the peer group are with a correlation coefficient of 0.64 strongly positively correlated. In line with Card, Heining, and Kline (2013), we also find a positive correlation between the worker and the firm-time and firm-occupation fixed effects of 0.14 and 0.16, respectively. These correlations illustrate the endog-enous sorting of high-ability workers into high-ability and high-wage peer groups and underscore the need to account for sorting in our estimates.
IV. Results A. Baseline Results
We report estimates for the impact of average peer quality, measured as the average worker fixed effect of coworkers in the peer group, on wages for the full sample in Table 3. Each column of the table introduces additional control variables to account for shared background characteristics. In column 1, we control for the worker’s own fixed effect ( a i in equation (4)), for quadratics in age and firm tenure (captured by x iojt in equation (4)), and for time-varying occupation fixed effects ( ω ot in equation (4)), in addition to firm fixed effects. The coefficient of 0.148 implies that a 10 percent increase in peer quality increases wages by 1.48 percent—an estimate roughly
0 0 0.05 0.1 −0.05 −0.1 50 Density 100 150 x
Peer FE change, stayers Peer FE change, movers
Figure 2. Variation of the Change in Peer Quality
Notes: The figure plots a kernel density estimate of the change in the average peer fixed effect (FE) separately for peer group stayers and peer group movers. Peer group quality varies more strongly for movers. For stayers, there is a mass point at zero, corresponding to stayers in peer groups that had no turnover. The figure is trimmed at the fifth percentile and ninety-fifth percentile of the distribution.
similar in magnitude to those reported by Lengermann (2002) and Battisti (2013) in a related specification. While this specification accounts for the possibility that work-ers employed in high-wage firms work with better pework-ers, it does not allow for firms which overpay specific occupations relative to the market to attract better workers into these occupations. To deal with this type of worker sorting, we control in column 2 for firm-occupation fixed effects ( θ oj in equation (4)) instead of simple firm fixed effects. This specification produces a much smaller estimate: a 10 percent increase in peer quality now increases the individual wage by only 0.66 percent. It does not yet filter out time-varying shocks at firm level common to all occupations in the firm. Such shocks turn out to be important: When adding time-varying firm fixed effects ( δ jt in equation (4)) in column 3, we find that a 10 percent increase in peer quality raises individual wages by merely 0.1 percent. Translated into standard deviations, this out-come implies that a one standard deviation increase in peer ability increases wages by 0.3 percentage points or 0.6 percent of a standard deviation. This effect is about 10–15 times smaller than the effects previously identified for productivity among supermar-ket cashiers in a single firm (Mas and Moretti 2009) and students carrying out a simple task in an experiment (Falk and Ichino 2006)—which incidentally are very close to the average effect reported by Herbst and Mas (2015) from a larger range of studies mostly covering specific field or lab settings. Therefore, we do not confirm similarly large spillover effects in wages for a representative set of occupations and firms.
B. Effects for Occupational Subgroups
Repetitive Occupations.—Even if peer effects in wages are small on average for a representative set of occupations, they might still be substantial for specific occupations. Hence, in panel A of Table 4, we report the results for the 5 percent
Table 3—Peer Effects in the Full Sample Observables,
occupation-year, and firm fixed effects
(1) Plus firm-occupation fixed effects (2) Plus firm-occupation and firm-year fixed
effects (3)
Average peer fixed effect 0.148 0.066 0.011
(0.002) (0.002) (0.001)
Worker fixed effects Yes Yes Yes
Occupation × year effects Yes Yes Yes
Firm effects Yes — —
Occupation × firm effects — Yes Yes
Firm × year effects — — Yes
Notes: The table shows the effect of average peer quality on the individual log wage in the overall sample. Peer quality is measured as the average fixed worker effect of the coworkers in the same three-digit occupation at the same firm in the same point of time. In column 1, we only control for worker fixed effects, firm fixed effects, occupation-by-year fixed effects, and quadratics in age and firm tenure. We then successively add firm-occupation fixed effects (column 2), and firm-by-year fixed effects (column 3). Specification (3) corresponds to the baseline specification described in equation (4) in the text. Coefficients can approximately be interpreted as elasticities, and the coefficient of 0.011 in the baseline specification in column 3 implies that a 10 percent increase in average peer quality increases wages by 0.1 percent. Bootstrapped standard errors with clustering at firm level in parentheses. N = 12,832,842.
of occupations with the most repetitive and predefined tasks, in which we expect particularly high peer pressure. These occupations also more closely resemble those used in earlier studies on peer pressure. The first three columns in the table refer to the baseline specification given by equation (4) and condition on occupation-year, firm-year, and firm-occupation fixed effects, meaning that they correspond to speci-fication (3) in the previous table.
For these repetitive occupations, we find a substantially larger effect of peer qual-ity on wages than in the full sample: a 10 percent increase in peer qualqual-ity raises wages by 0.64 percent (see column 1) compared to the effect of 0.1 percent in the full sample (see column 3 of Table 3). This outcome implies that a one standard devi-ation increase in peer quality increases the wage by 0.84 percent, about one-half the size of the peer effects in coworker productivity identified by Mas and Moretti (2009) and Falk and Ichino (2006) and in the meta-analysis by Herbst and Mas (2015).
Table 4—Peer Effects in Subsamples of Occupations Panel A. Peer effects for subsamples of low-skilled occupations
Baseline specification of equation (4)
Within–peer group estimator of equation (6) 5% most repetitive occupations As in case studies Low learning content 5% most repetitive occupations (1) (2) (3) (4)
Average peer fixed effect 0.064 0.067 0.052 0.061
(0.0070) (0.0116) (0.0031) (0.006) Panel B. Peer effects for subsamples of high-skilled occupations
Baseline specification of equation (4)
Within–peer group estimator of equation (6) 10% most skilled occupations 10% most innovative occupations High learning content 10% most skilled occupations (1) (2) (3) (4)
Average peer fixed effect 0.013 0.007 0.017 0.016
(0.0039) (0.0044) (0.0028) (0.004)
Notes: The first three columns of the table replicate the baseline peer effects estimates of column 3 in Table 3 for different occupational groups. See online Appendix Table F.3 for a full list of occupations in each of the subsamples used in this table, and Section IIIC in the text for a description of the way in which the different subsamples were constructed. In panel A, column 1, we show the effect for the 5 percent most repetitive occupations. In panel A, col-umn 2, we show the effect for agricultural helpers, cashiers and data entry workers, which have been used in related case-studies on peer effects in the workplace. In panel A, column 3, we report the effect for occupations character-ized by standardcharacter-ized tasks (as the 5 percent most repetitive occupations) and limited learning content (i.e., cashiers, warehouse workers, drivers, removal workers, cleaners, agricultural helpers, and waiters). In panel B, column 1 we present results for the 10 percent most skilled occupations, as measured by the share of workers with a college degree in that occupation. In panel B, column 2 we present results for the 10 percent most innovative occupations, defined by occupation averages of workers’ responses to an index of how frequently they need to experiment with new ideas. In panel B, column 3 we present results for occupations with complex tasks and a high learning content (such as doctors, lawyers, scientists, teachers, and academics). In column 4, we present the within–peer group esti-mate for, as in column 1, the 5 percent most repetitive and 10 percent most skilled occupations, see equation (6) in the text. The within-estimator is based on pre-estimated worker fixed effects from the baseline model in equation (4) in the text. Bootstrapped standard errors with clustering at firm level in parentheses. N = 12,832,842.
Column 2 of panel A shows the peer effect for the three occupations used in earlier studies (agricultural helpers, cashiers, and data entry workers), which is remarkably similar in magnitude to that for the 5 percent most repetitive occupations shown in column 1. Column 3 reports the results for the handpicked group of occupations in which we expect easily observable output and, following initial induction, lim-ited on-the-job learning. The estimated effect for this occupational group is slightly smaller than that for the 5 percent most repetitive occupations sample but still about five times as large as the effect for the full occupational sample.
Column 4 reports estimates using the within–peer group estimator for the 5 per-cent most repetitive occupations (see equation (6)). As we point out above, this estimator is robust to unobserved time-varying peer group wage shocks that are correlated with shocks to true or estimated peer group quality. The estimated peer effects based on the within–peer group specification is very close to the effect derived in the respective baseline specification. This similarity in estimates corroborates that time-varying peer group-specific wage shocks are not important, and provides reas-surance that we are picking up a true peer effect rather than a spurious correlation.
High-Skilled Occupations.—In panel B of Table 4, we restrict the analysis to particularly high-skilled and innovative occupations with a high scope for learning, in which we expect knowledge spillover to be important. Yet regardless of how we define high-skilled occupations (columns 1 to 3), and whether or not we exploit variation in peer ability within peer groups only (column 4), peer effects in these groups are small and resemble those in the full sample. Overall, therefore, we iden-tify sizeable peer effects in wages only in occupations characterized by standard-ized tasks and low learning content, which are exactly the occupations in which we expect peer pressure to matter and which closely resemble the specific occupations investigated in the extant studies on peer pressure.
By looking at the 5 percent most repetitive and the 10 percent most skilled occu-pations we have distinguished between the two extreme ends of the two indexes of repetitiveness and skill from which the definition of these groups was derived. In Figure 3 we show results from a more complete analysis that lets the peer effect coefficient vary by bins of these two indexes. They show a symmetric pattern, with highest peer effects in the most repetitive/least skilled categories, smallest peer effects in the middle categories, and again slightly higher but still small effects in the least repetitive/most skilled categories.14 The U-shape of the estimated peer
effects in these indexes provides support for our hypothesis that peer pressure and knowledge spillover are two possible mechanisms for peer effects, where the former operates predominantly in the most repetitive (and least skilled) occupations, while the latter is most pronounced in the least repetitive and most skilled occupations.
C. Timing of Effects
Figure 4 provides a first visual impression of the timing of the wage response to a change in peer quality in the 5 percent most repetitive occupations where peer
